MAALOUFimad02/Machine_Learning_Training
0
1<!DOCTYPE html>2<html lang="fr">3<head>4 <meta charset="UTF-8">5 <title>Cours de Machine Learning - ML Academy</title>6 <link rel="preconnect" href="https://fonts.googleapis.com">7 <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>8 <link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&family=JetBrains+Mono:wght@400;500&display=swap" rel="stylesheet">9 <script>10 window.MathJax = {11 tex: { inlineMath: [['$', '$']], displayMath: [['$$', '$$']] },12 options: { skipHtmlTags: ['script','noscript','style','textarea','pre'] }13 };14 </script>15 <script src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js"></script>16 <style>17 /* Reset */18 * { box-sizing: border-box; margin: 0; padding: 0; }19 body { margin: 0; padding: 0; font-family: 'Inter', -apple-system, sans-serif; }20 21 /* Page setup */22 @page {23 size: A4;24 margin: 2.5cm 2cm;25 @bottom-center {26 content: "Imad Maalouf — ML Academy";27 font-size: 9pt;28 color: #666;29 }30 @bottom-right {31 content: counter(page);32 font-size: 9pt;33 color: #666;34 }35 }36 37 @page :first {38 margin: 0;39 @bottom-center { content: none; }40 @bottom-right { content: none; }41 }42 43 /* Cover page */44 .cover {45 width: 210mm;46 height: 297mm;47 margin: 0;48 position: relative;49 overflow: hidden;50 page-break-after: always;51 background: linear-gradient(135deg, #0f0f1a 0%, #1a1a2e 50%, #252542 100%);52 }53 54 .cover-decoration {55 position: absolute;56 top: 0;57 left: 0;58 right: 0;59 bottom: 0;60 overflow: hidden;61 }62 63 .cover-circle {64 position: absolute;65 border-radius: 50%;66 opacity: 0.1;67 }68 69 .cover-circle-1 {70 width: 400px;71 height: 400px;72 background: #6366f1;73 top: -100px;74 right: -100px;75 }76 77 .cover-circle-2 {78 width: 300px;79 height: 300px;80 background: #06b6d4;81 bottom: 10%;82 left: -80px;83 }84 85 .cover-circle-3 {86 width: 200px;87 height: 200px;88 background: #f59e0b;89 top: 40%;90 right: 15%;91 }92 93 .cover-content {94 position: absolute;95 top: 50%;96 left: 50%;97 transform: translate(-50%, -50%);98 text-align: center;99 width: 80%;100 z-index: 1;101 }102 103 .cover-badge {104 display: inline-block;105 padding: 8px 16px;106 background: rgba(99, 102, 241, 0.2);107 border: 1px solid rgba(99, 102, 241, 0.4);108 border-radius: 20px;109 font-family: 'JetBrains Mono', monospace;110 font-size: 10pt;111 color: #818cf8;112 text-transform: uppercase;113 letter-spacing: 2px;114 margin-bottom: 30px;115 }116 117 .cover-title {118 font-size: 36pt;119 font-weight: 700;120 color: #f8fafc;121 margin-bottom: 15px;122 line-height: 1.2;123 }124 125 .cover-subtitle {126 font-size: 14pt;127 color: #94a3b8;128 margin-bottom: 60px;129 }130 131 .cover-author {132 font-size: 12pt;133 color: #818cf8;134 margin-bottom: 10px;135 }136 137 .cover-info {138 font-size: 10pt;139 color: #64748b;140 }141 142 /* Content styles */143 .content {144 padding: 0;145 }146 147 h1 {148 font-size: 20pt;149 font-weight: 700;150 color: #1a1a2e;151 margin: 30pt 0 15pt 0;152 padding-bottom: 8pt;153 border-bottom: 2px solid #6366f1;154 page-break-after: avoid;155 }156 157 h2 {158 font-size: 14pt;159 font-weight: 600;160 color: #1a1a2e;161 margin: 20pt 0 10pt 0;162 page-break-after: avoid;163 }164 165 h3 {166 font-size: 12pt;167 font-weight: 600;168 color: #333;169 margin: 15pt 0 8pt 0;170 page-break-after: avoid;171 }172 173 p {174 font-size: 11pt;175 line-height: 1.7;176 color: #333;177 margin-bottom: 10pt;178 text-align: justify;179 text-align-last: left;180 }181 182 strong {183 font-weight: 600;184 color: #1a1a2e;185 }186 187 em {188 font-style: italic;189 color: #444;190 }191 192 /* Info box */193 .info-box {194 background: #f8f9fa;195 border-left: 4px solid #6366f1;196 padding: 12pt 15pt;197 margin: 15pt 0;198 page-break-inside: avoid;199 }200 201 .info-box-title {202 font-weight: 600;203 color: #6366f1;204 margin-bottom: 5pt;205 font-size: 10pt;206 }207 208 .info-box p {209 margin: 0;210 font-size: 10pt;211 }212 213 /* Equation block */214 .equation-block {215 background: #f8f9fa;216 border: 1px solid #e2e8f0;217 border-left: 3px solid #6366f1;218 border-radius: 4px;219 padding: 15pt;220 margin: 15pt 0;221 text-align: center;222 page-break-inside: avoid;223 }224 225 .equation-label {226 display: block;227 font-family: 'JetBrains Mono', monospace;228 font-size: 8pt;229 color: #64748b;230 text-align: right;231 margin-top: 8pt;232 }233 234 /* Cards grid */235 .cards-grid {236 display: grid;237 grid-template-columns: repeat(3, 1fr);238 gap: 12pt;239 margin: 15pt 0;240 }241 242 .card {243 background: #f8f9fa;244 border: 1px solid #e2e8f0;245 border-radius: 6px;246 padding: 12pt;247 page-break-inside: avoid;248 }249 250 .card-icon {251 width: 32px;252 height: 32px;253 background: linear-gradient(135deg, #6366f1, #06b6d4);254 border-radius: 6px;255 display: flex;256 align-items: center;257 justify-content: center;258 font-size: 12pt;259 font-weight: 700;260 color: white;261 margin-bottom: 8pt;262 }263 264 .card-title {265 font-size: 10pt;266 font-weight: 600;267 color: #1a1a2e;268 margin-bottom: 5pt;269 }270 271 .card-text {272 font-size: 9pt;273 color: #555;274 line-height: 1.5;275 }276 277 /* Pipeline */278 .pipeline {279 display: flex;280 gap: 8pt;281 margin: 15pt 0;282 padding: 12pt;283 background: #f8f9fa;284 border: 1px solid #e2e8f0;285 border-radius: 6px;286 }287 288 .pipeline-step {289 flex: 1;290 text-align: center;291 position: relative;292 }293 294 .pipeline-step:not(:last-child)::after {295 content: '->';296 position: absolute;297 right: -10pt;298 top: 50%;299 transform: translateY(-50%);300 color: #6366f1;301 font-family: 'JetBrains Mono', monospace;302 font-size: 10pt;303 }304 305 .pipeline-num {306 width: 24pt;307 height: 24pt;308 background: linear-gradient(135deg, #6366f1, #06b6d4);309 border-radius: 50%;310 display: flex;311 align-items: center;312 justify-content: center;313 font-family: 'JetBrains Mono', monospace;314 font-size: 9pt;315 font-weight: 700;316 color: white;317 margin: 0 auto 5pt;318 }319 320 .pipeline-label {321 font-size: 8pt;322 font-weight: 600;323 color: #1a1a2e;324 }325 326 .pipeline-desc {327 font-size: 7pt;328 color: #64748b;329 }330 331 /* Comparison */332 .comparison-grid {333 display: grid;334 grid-template-columns: 1fr 1fr;335 gap: 12pt;336 margin: 15pt 0;337 }338 339 .comparison-box {340 background: #f8f9fa;341 border: 1px solid #e2e8f0;342 border-radius: 6px;343 padding: 12pt;344 page-break-inside: avoid;345 }346 347 .comparison-box h4 {348 font-size: 10pt;349 font-weight: 600;350 color: #1a1a2e;351 margin-bottom: 8pt;352 padding-bottom: 5pt;353 border-bottom: 1px solid #e2e8f0;354 }355 356 .comparison-box ul {357 list-style: none;358 padding: 0;359 margin: 0;360 }361 362 .comparison-box li {363 padding: 3pt 0;364 font-size: 9pt;365 color: #444;366 }367 368 .comparison-box li::before {369 content: '>';370 color: #6366f1;371 margin-right: 6pt;372 font-family: 'JetBrains Mono', monospace;373 }374 375 /* Metric cards */376 .metric-row {377 display: grid;378 grid-template-columns: repeat(3, 1fr);379 gap: 12pt;380 margin: 15pt 0;381 }382 383 .metric-card {384 background: #f8f9fa;385 border: 1px solid #e2e8f0;386 border-radius: 6px;387 padding: 12pt;388 text-align: center;389 page-break-inside: avoid;390 }391 392 .metric-name {393 font-family: 'JetBrains Mono', monospace;394 font-size: 7pt;395 color: #64748b;396 text-transform: uppercase;397 letter-spacing: 1px;398 margin-bottom: 5pt;399 }400 401 .metric-formula {402 font-family: 'JetBrains Mono', monospace;403 font-size: 9pt;404 color: #6366f1;405 font-weight: 600;406 margin-bottom: 3pt;407 }408 409 .metric-desc {410 font-size: 7pt;411 color: #64748b;412 }413 414 /* Table */415 table {416 width: 100%;417 border-collapse: collapse;418 margin: 15pt 0;419 font-size: 9pt;420 page-break-inside: avoid;421 }422 423 thead {424 display: table-header-group;425 }426 427 th {428 background: #f1f5f9;429 color: #475569;430 font-family: 'JetBrains Mono', monospace;431 font-size: 8pt;432 font-weight: 600;433 text-transform: uppercase;434 letter-spacing: 0.5px;435 padding: 8pt 10pt;436 text-align: left;437 border-top: 2px solid #333;438 border-bottom: 1px solid #333;439 }440 441 td {442 padding: 8pt 10pt;443 border-bottom: 1px solid #e2e8f0;444 color: #333;445 }446 447 tbody tr:last-child td {448 border-bottom: 2px solid #333;449 }450 451 /* Page break utilities */452 .page-break {453 page-break-after: always;454 }455 456 /* Prevent overflow */457 pre, table, figure, img, svg, blockquote {458 max-width: 100%;459 box-sizing: border-box;460 }461 462 a { word-break: break-all; }463 464 /* Author footer on each page */465 .author-footer {466 margin-top: 30pt;467 padding-top: 15pt;468 border-top: 1px solid #e2e8f0;469 text-align: center;470 }471 472 .author-name {473 font-size: 10pt;474 font-weight: 600;475 color: #6366f1;476 }477 478 .author-contact {479 font-size: 8pt;480 color: #64748b;481 margin-top: 3pt;482 }483 </style>484</head>485<body>486 <!-- Cover Page -->487 <div class="cover">488 <div class="cover-decoration">489 <div class="cover-circle cover-circle-1"></div>490 <div class="cover-circle cover-circle-2"></div>491 <div class="cover-circle cover-circle-3"></div>492 </div>493 <div class="cover-content">494 <div class="cover-badge">Formation 2025/2026</div>495 <h1 class="cover-title">Cours de<br>Machine Learning</h1>496 <p class="cover-subtitle">De la theorie a la pratique — Algorithmes fondamentaux et techniques avancees</p>497 <p class="cover-author">Formateur : Imad Maalouf</p>498 <p class="cover-info">ML Academy — GE-MCI 4A</p>499 </div>500 </div>501 502 <!-- Content -->503 <div class="content">504 <h1>1. Introduction au Machine Learning</h1>505 506 <p>507 Le <strong>Machine Learning (ML)</strong> est une branche de l'intelligence artificielle 508 qui permet aux machines d'apprendre a partir de donnees <em>sans etre explicitement programmees</em> 509 pour chaque tache. Au lieu de coder des regles a la main, on montre des exemples au modele 510 et il decouvre lui-meme les patterns.511 </p>512 513 <div class="info-box">514 <div class="info-box-title">Idee fondamentale</div>515 <p>516 On cherche a approximer une fonction inconnue $f$ telle que $\hat{y} = f(x_1, x_2, \ldots, x_n)$. 517 Le modele ML apprend cette fonction a partir d'exemples $(x, y)$ connus.518 </p>519 </div>520 521 <h2>1.1 Types d'apprentissage</h2>522 523 <div class="cards-grid">524 <div class="card">525 <div class="card-icon">S</div>526 <div class="card-title">Supervise</div>527 <div class="card-text">Donnees labellisees $(x, y)$ : regression et classification.</div>528 </div>529 <div class="card">530 <div class="card-icon">N</div>531 <div class="card-title">Non supervise</div>532 <div class="card-text">Pas de labels : clustering, reduction de dimension.</div>533 </div>534 <div class="card">535 <div class="card-icon">R</div>536 <div class="card-title">Par renforcement</div>537 <div class="card-text">Agent apprend via actions-recompenses.</div>538 </div>539 </div>540 541 <h2>1.2 Pipeline ML typique</h2>542 543 <div class="pipeline">544 <div class="pipeline-step">545 <div class="pipeline-num">1</div>546 <div class="pipeline-label">Donnees</div>547 <div class="pipeline-desc">Collecte & nettoyage</div>548 </div>549 <div class="pipeline-step">550 <div class="pipeline-num">2</div>551 <div class="pipeline-label">Features</div>552 <div class="pipeline-desc">Engineering</div>553 </div>554 <div class="pipeline-step">555 <div class="pipeline-num">3</div>556 <div class="pipeline-label">Split</div>557 <div class="pipeline-desc">Train / Test</div>558 </div>559 <div class="pipeline-step">560 <div class="pipeline-num">4</div>561 <div class="pipeline-label">Modele</div>562 <div class="pipeline-desc">Entrainement</div>563 </div>564 <div class="pipeline-step">565 <div class="pipeline-num">5</div>566 <div class="pipeline-label">Evaluation</div>567 <div class="pipeline-desc">Metriques</div>568 </div>569 <div class="pipeline-step">570 <div class="pipeline-num">6</div>571 <div class="pipeline-label">Production</div>572 <div class="pipeline-desc">Deploiement</div>573 </div>574 </div>575 576 <div class="page-break"></div>577 578 <h1>2. Regression Lineaire</h1>579 580 <p>581 La <strong>regression lineaire</strong> modelise la relation entre les features et la cible 582 par une fonction affine. C'est l'algorithme le plus simple mais souvent tres efficace 583 comme baseline.584 </p>585 586 <div class="equation-block">587 $$\hat{y} = w_0 + w_1 x_1 + w_2 x_2 + \cdots + w_n x_n = \mathbf{w}^T \mathbf{x}$$588 <span class="equation-label">Modele lineaire avec coefficients $\mathbf{w}$</span>589 </div>590 591 <p>592 L'objectif est de minimiser l'erreur quadratique moyenne (MSE) :593 </p>594 595 <div class="equation-block">596 $$\text{MSE} = \frac{1}{m} \sum_{i=1}^{m} (y_i - \hat{y}_i)^2$$597 <span class="equation-label">Fonction de cout : moyenne des erreurs quadratiques</span>598 </div>599 600 <div class="info-box">601 <div class="info-box-title">Solution analytique</div>602 <p>603 La regression lineaire admet une solution fermee : 604 $\mathbf{w}^* = (X^T X)^{-1} X^T Y$. Pas besoin d'iterations !605 </p>606 </div>607 608 <h2>2.1 Avantages et limitations</h2>609 610 <div class="comparison-grid">611 <div class="comparison-box">612 <h4>[+] Avantages</h4>613 <ul>614 <li>Tres rapide a entrainer</li>615 <li>Interpretable (coefficients)</li>616 <li>Pas d'hyperparametres</li>617 <li>Excellent baseline</li>618 </ul>619 </div>620 <div class="comparison-box">621 <h4>[-] Limitations</h4>622 <ul>623 <li>Relation lineaire uniquement</li>624 <li>Sensible aux outliers</li>625 <li>Performance decroit en haute dimension</li>626 </ul>627 </div>628 </div>629 630 <div class="page-break"></div>631 632 <h1>3. Regression Logistique</h1>633 634 <p>635 Malgre son nom, la <strong>regression logistique</strong> est un algorithme de <em>classification</em>. 636 Elle predit la probabilite d'appartenance a une classe en utilisant la fonction sigmoide.637 </p>638 639 <div class="equation-block">640 $$P(y=1|\mathbf{x}) = \sigma(\mathbf{w}^T \mathbf{x}) = \frac{1}{1 + e^{-\mathbf{w}^T \mathbf{x}}}$$641 <span class="equation-label">Fonction sigmoide pour la classification binaire</span>642 </div>643 644 <div class="info-box">645 <div class="info-box-title">Cas d'usage : Dataset Titanic</div>646 <p>647 Predire la survie des passagers du Titanic a partir de leur age, sexe, 648 classe de billet, etc. Un classique du ML pour debuter !649 </p>650 </div>651 652 <div class="page-break"></div>653 654 <h1>4. Random Forest</h1>655 656 <p>657 <strong>Random Forest</strong> est un ensemble d'arbres de decision qui votent pour predire. 658 C'est l'un des algorithmes les plus populaires en ML applique : performant, robuste, 659 peu sensible au tuning.660 </p>661 662 <h2>4.1 Algorithme : Bagging + Random Splits</h2>663 664 <div class="pipeline">665 <div class="pipeline-step">666 <div class="pipeline-num">1</div>667 <div class="pipeline-label">Bootstrap</div>668 <div class="pipeline-desc">Echantillons aleatoires</div>669 </div>670 <div class="pipeline-step">671 <div class="pipeline-num">2</div>672 <div class="pipeline-label">Splits</div>673 <div class="pipeline-desc">Features aleatoires</div>674 </div>675 <div class="pipeline-step">676 <div class="pipeline-num">3</div>677 <div class="pipeline-label">Arbres</div>678 <div class="pipeline-desc">N arbres independants</div>679 </div>680 <div class="pipeline-step">681 <div class="pipeline-num">4</div>682 <div class="pipeline-label">Vote</div>683 <div class="pipeline-desc">Moyenne ou mode</div>684 </div>685 </div>686 687 <h2>4.2 Hyperparametres cles</h2>688 689 <div class="metric-row">690 <div class="metric-card">691 <div class="metric-name">n_estimators</div>692 <div class="metric-formula">100 - 500</div>693 <div class="metric-desc">Nombre d'arbres</div>694 </div>695 <div class="metric-card">696 <div class="metric-name">max_depth</div>697 <div class="metric-formula">10 - 30</div>698 <div class="metric-desc">Profondeur max</div>699 </div>700 <div class="metric-card">701 <div class="metric-name">min_samples_split</div>702 <div class="metric-formula">2 - 10</div>703 <div class="metric-desc">Min pour splitter</div>704 </div>705 </div>706 707 <div class="info-box">708 <div class="info-box-title">Feature Importance</div>709 <p>710 Random Forest fournit automatiquement l'importance de chaque feature, 711 ce qui aide a comprendre quelles variables influencent le plus les predictions.712 </p>713 </div>714 715 <div class="page-break"></div>716 717 <h1>5. Reseaux de Neurones</h1>718 719 <p>720 Les <strong>reseaux de neurones</strong> sont inspires du cerveau humain : des couches de neurones 721 interconnectes executent des transformations non-lineaires. Ils excellent pour les patterns complexes.722 </p>723 724 <h2>5.1 Fonctionnement : Forward + Backprop</h2>725 726 <div class="cards-grid">727 <div class="card">728 <div class="card-icon">F</div>729 <div class="card-title">Forward Pass</div>730 <div class="card-text">Donnees traversent les couches : $\mathbf{h}_1 = \sigma(W_1 \mathbf{x} + b_1)$</div>731 </div>732 <div class="card">733 <div class="card-icon">L</div>734 <div class="card-title">Loss Computation</div>735 <div class="card-text">Compare prediction vs realite : $L = \frac{1}{m} \sum (y - \hat{y})^2$</div>736 </div>737 <div class="card">738 <div class="card-icon">B</div>739 <div class="card-title">Backpropagation</div>740 <div class="card-text">Calcule les gradients via la chaine de derivation</div>741 </div>742 </div>743 744 <h2>5.2 Fonctions d'activation</h2>745 746 <div class="metric-row">747 <div class="metric-card">748 <div class="metric-name">ReLU</div>749 <div class="metric-formula">$f(x) = \max(0, x)$</div>750 <div class="metric-desc">Couches cachees</div>751 </div>752 <div class="metric-card">753 <div class="metric-name">Sigmoid</div>754 <div class="metric-formula">$f(x) = \frac{1}{1 + e^{-x}}$</div>755 <div class="metric-desc">Classification binaire</div>756 </div>757 <div class="metric-card">758 <div class="metric-name">Softmax</div>759 <div class="metric-formula">$f(x_i) = \frac{e^{x_i}}{\sum_j e^{x_j}}$</div>760 <div class="metric-desc">Classification multi-classe</div>761 </div>762 </div>763 764 <div class="page-break"></div>765 766 <h1>6. LSTM et Series Temporelles</h1>767 768 <p>769 <strong>LSTM (Long Short-Term Memory)</strong> est un type de reseau neuronal pour series temporelles. 770 Il peut "retenir" l'information sur de longues periodes — crucial pour les predictions temporelles.771 </p>772 773 <div class="info-box">774 <div class="info-box-title">Probleme des RNN vanilla</div>775 <p>776 Les gradients disparaissent (vanishing) ou explosent (exploding) sur de longues sequences. 777 Le LSTM resout ce probleme avec son <strong>cell state</strong>.778 </p>779 </div>780 781 <h2>6.1 Les trois portes du LSTM</h2>782 783 <div class="metric-row">784 <div class="metric-card">785 <div class="metric-name">Forget Gate</div>786 <div class="metric-formula">$f_t = \sigma(W_f [h_{t-1}, x_t] + b_f)$</div>787 <div class="metric-desc">Quoi oublier ?</div>788 </div>789 <div class="metric-card">790 <div class="metric-name">Input Gate</div>791 <div class="metric-formula">$i_t = \sigma(W_i [h_{t-1}, x_t] + b_i)$</div>792 <div class="metric-desc">Quoi ajouter ?</div>793 </div>794 <div class="metric-card">795 <div class="metric-name">Output Gate</div>796 <div class="metric-formula">$o_t = \sigma(W_o [h_{t-1}, x_t] + b_o)$</div>797 <div class="metric-desc">Quoi exposer ?</div>798 </div>799 </div>800 801 <div class="info-box">802 <div class="info-box-title">Cas d'usage</div>803 <p>804 Prediction de prix boursiers, meteo, consommation energetique, 805 traitement du langage naturel (NLP)...806 </p>807 </div>808 809 <div class="page-break"></div>810 811 <h1>7. Metriques de Performance</h1>812 813 <p>814 Evaluer correctement un modele est crucial. Les bonnes metriques dependent du type de probleme 815 (regression vs classification) et des objectifs metier.816 </p>817 818 <h2>7.1 Regression</h2>819 820 <div class="metric-row">821 <div class="metric-card">822 <div class="metric-name">MAE</div>823 <div class="metric-formula">$\frac{1}{n}\sum|y_i - \hat{y}_i|$</div>824 <div class="metric-desc">Robuste aux outliers</div>825 </div>826 <div class="metric-card">827 <div class="metric-name">RMSE</div>828 <div class="metric-formula">$\sqrt{\frac{1}{n}\sum(y_i - \hat{y}_i)^2}$</div>829 <div class="metric-desc">Penalise les grandes erreurs</div>830 </div>831 <div class="metric-card">832 <div class="metric-name">R2</div>833 <div class="metric-formula">$1 - \frac{SS_{res}}{SS_{tot}}$</div>834 <div class="metric-desc">% variance expliquee</div>835 </div>836 </div>837 838 <h2>7.2 Classification</h2>839 840 <table>841 <thead>842 <tr>843 <th>Metrique</th>844 <th>Formule</th>845 <th>Usage</th>846 </tr>847 </thead>848 <tbody>849 <tr>850 <td><strong>Accuracy</strong></td>851 <td>$(TP + TN) / Total$</td>852 <td>Classes equilibrees</td>853 </tr>854 <tr>855 <td><strong>Precision</strong></td>856 <td>$TP / (TP + FP)$</td>857 <td>Minimiser faux positifs</td>858 </tr>859 <tr>860 <td><strong>Recall</strong></td>861 <td>$TP / (TP + FN)$</td>862 <td>Minimiser faux negatifs</td>863 </tr>864 <tr>865 <td><strong>F1-Score</strong></td>866 <td>$2 \cdot \frac{P \cdot R}{P + R}$</td>867 <td>Classes desequilibrees</td>868 </tr>869 </tbody>870 </table>871 872 <div class="page-break"></div>873 874 <h1>8. Optimisation et Regularisation</h1>875 876 <p>877 Pour eviter le <strong>surapprentissage (overfitting)</strong> et ameliorer la generalisation, 878 plusieurs techniques existent.879 </p>880 881 <div class="cards-grid">882 <div class="card">883 <div class="card-icon">D</div>884 <div class="card-title">Dropout</div>885 <div class="card-text">Desactive aleatoirement des neurones pendant l'entrainement.</div>886 </div>887 <div class="card">888 <div class="card-icon">E</div>889 <div class="card-title">Early Stopping</div>890 <div class="card-text">Arrete l'entrainement quand la validation stagne.</div>891 </div>892 <div class="card">893 <div class="card-icon">L</div>894 <div class="card-title">L2 Regularization</div>895 <div class="card-text">Penalise les grands poids : $L_{total} = L_{data} + \lambda \sum w^2$</div>896 </div>897 </div>898 899 <div class="info-box">900 <div class="info-box-title">Regle d'or</div>901 <p>902 Toujours comparer les metriques sur <strong>train</strong> ET <strong>test</strong>. 903 Un grand ecart = overfitting. Objectif : R2 train ≈ R2 test.904 </p>905 </div>906 907 <div class="author-footer">908 <div class="author-name">Imad Maalouf</div>909 <div class="author-contact">910 imadmaalouf02@gmail.com | github.com/imadmaalouf02 | huggingface.co/spaces/MAALOOUF/ML_Training911 </div>912 </div>913 </div>914</body>915</html>916